What problem does it solve? Bringing up a new CUDA-compatible accelerator (MetaX, Hygon DCU, PPU) normally means writing hundreds of operator kernels. This Skill shows how to avoid that entirely by reusing the vendor's own CUDA-shaped torch build: extract its libtorch_cuda.so, preload it before importing torch, and box flagos tensors to the CUDA dispatch key with zero data copies. ## Core Features & Use Cases - Compatibility verification: Mechanically prove a vendor ships a usable libtorch_cuda.so by inspecting ATen symbols and dumping the dispatcher table before committing to this path. - Vendor .so extraction and wheel bundling: Download (never pip-install) the vendor torch wheel, stage libtorch_cuda.so and libc10_cuda.so into .libtorch_cuda_assets, and let setup.py bundle them into the built wheel. - Load-timing and build wiring: Enforce the mandatory LD_PRELOAD-before-import-torch constraint via scripts/vendor/with_cuda_libtorch.sh, and wire the correct FLAGOS_ACCELERATOR branch (DCU-style glob, PPU-style detection, or MetaX-style shim) with a g++-only, no-nvcc build. - Use Case: You receive a Hygon DCU machine whose DTK ships a CUDA-compatible torch. Follow the steps to confirm dispatcher entries for mm/add/_softmax/bmm, bundle the extracted libraries, and run the operator integration suite with tensors staying on the flagos device. ## Quick Start Verify the vendor's libtorch_cuda.so registers CUDA kernels for aten::mm, then extract it from the vendor wheel and run the torch_fl operator tests through the with_cuda_libtorch.sh preload wrapper.
